- ML for security and security for ML
Trainer Names: Nikhil Joshi
Title: ML for security and security for ML
Duration: 4 Days
Dates: 1st - 4th March 2021
Time: 10.00 AM to 2.00 PM IST
Delivery Mode: via Zoom.us and Discord Platform
Machine learning / Deep learning is under exponential growth these days. Businesses, Academia and tech enthusiasts are really hyped about trying out Deep learning to solve their problems. A lot of students, professionals and researchers are driven to learn this new cool tech. Just like every other technology, ML comes with awesome applications topped with some serious implications.
So, join the 4 days online expedition specially designed for security professionals to understand, build and hack Machine Learning applications. The course is divided into two parts, ML4SEC & SEC4ML. ML4SEC will focus on nitty-gritties of building ML applications. Then learn to hack them in SEC4ML part.
Considering no prior knowledge of mathematics and ML, we will try to build the intuition behind algorithms. Attendees will go through hands-on experience in building ML powered defensive and offensive security tools. In-depth understanding of the entire ML pipeline is provided. Which consists of pre-processing data, building ML models, training and evaluating them and using trained models for prediction. Well known machine learning libraries like Tensorflow, Keras, Pytorch, sklearn, etc. will be used. At the end you will be ready with end-to-end and ready to apply ML Gyan for security professionals.
This part will address the vulnerabilities (like Adversarial learning, Model stealing, Data poisoning, Model Inference, etc) in state of the art machine learning methodologies. Lab material will consist of Vulnerable Machine Learning applications that can be exploited to provide a thorough understanding of discussed vulnerabilities. Possible mitigation to these vulnerabilities will also be discussed.
DAY 1 and 2:
In this session, we will build up our understanding of basic yet state of the art machine learning algorithms. Discuss mathemagic behind why these models work the way they do. Build some smart Machine Learning applications and evaluate them. By the end, we will get an idea of how to solve a real-world problem using machine learning.
- Introduction to Machine learning
- Common use cases, where to use and where not to use machine learning
- Introduction to different python libraries/packages like keras, tensorflow, sklearn
- Overview of how machine learning models are built and deployed in production
- Understanding Mathematics and intuition behind used machine learning algorithms
- Supervised learning
- Linear regression, logistic regression, Neural nets and similar classifiers
- Unsupervised learning
- Clustering algorithms like k-means
- Semi-supervised learning
- Supervised learning
- Brief introduction on data pre-processing with demo
- Cooking a dataset so that it can be consumed by discussed models
- Feature engineering: Decreasing the dimensionality of problem or adding more features to dataset
- Removing unnecessary data and handling different data types
- Dealing with incomplete data
- Applications of machine learning in security domain with hands on examples
- Detailed process of how to leverage previously discussed knowledge to build applications in defensive as well as offensive security.
- Image classifier using deep learning
- Defensive sec:
- Web access firewalls
- Intrusion detection systems
- Malware detection engine
- Offensive sec
- Machine learning for phishing
- Machine learning for fuzzing
- Evaluate the built models using different evaluation parameters.
- Now that we have made our systems “Intelligent”, is it possible to fool them? Are these applications hackable
DAY 3 and 4:
SEC4ML In this session we will have a deeper look on different flaws in how ML/DL algorithms are implemented. Hands on examples explaining and attacking such vulnerable implementations. Also, discussion on possible mitigation.
- Brief introduction to vulnerabilities in Machine Learning
- Discussion on various ways of compromising machine learning apps
- Adversarial learning Attacks
- Introduction and mathematical intuition behind the existence of this flaw
- Demo and hands on practice of fooling very accurate state-of-the art Image classifiers
- Analysing why this attack works
- Possible mitigation
- Model stealing Attacks
- How proprietary ML models can be stolen by attacker, making him/her to use the models for FREE
- Stealing offline ML models that are deployed on device with installer packages
- Stealing models that are deployed on cloud with restricted access via APIs
- Model Skewing and data poisoning attacks
- How and why this attack works
- Hands on example of bypassing ML based 99.99% accurate Spam Filters
- Possible Mitigation
- Some miscellaneous attacks
- Discussion on other lesser addressed vulnerabilities and real world impact.
- CTF challenge focusing on one of the discussed vulnerabilities
What to bring
- Laptop with 8GB+ RAM
- 20 GB space
- Virtual box (latest version)
- Any flavour of Linux is preferred over windows
- Open mind made up for some intense mathemagic
- Basic knowledge of python is good to have but not required
- Basic of Linux and Virtualbox
Who should attend
- Machine learning enthusiasts and professionals
- Security researchers and pentesters looking forward to implement ML/DL in their research
- Pentesters willing to explore new ways to pentest Machine learning applications
- Students with computer science background and a taste for ML and infosec
- Course slides and notes.
- Precooked VM ready to run lab and exercise codes for post training practice.
What to expect
- Thorough understanding of basic machine learning methodologies
- Hands on practice on Specially crafted labs for ML and Infosec enthusiasts
- End-to-end and ready to apply ML knowledge for security professionals
- Good understanding of Machine learning vulnerabilities
- Hands on experience with well known machine learning libraries
- Lab material for post-course practice
What not to expect
- Being a ML pro in two days
- Heavy mathematical background of Machine Learning concepts
Nikhil Joshi is an AI Security Researcher at Payatu. He has been the Machine Learning guy for more than 4 years and currently working on implementations of ML in offensive and defensive security products. At Payatu, He has orchestrated methodologies to pen-test Machine Learning application against ML specific vulnerabilities and loves to explore new ways to hack ML powered applications. Parallelly Nikhil's research is focused on security implications in Deep Learning applications such as Adversarial Learning, Model stealing attacks, Data poisoning, etc.
Nikhil is an active member of local Data Science and Security groups and has delivered multiple talks and workshops. Also has spoken at HITB Amsterdam, PhDays Russia and presented his research at IEEE conference. He is a trainer at Nullcon and troopers. Being an Applied Mathematics enthusiast, recent advances in Machine Learning and its applications in security, behavioural science and telecom are of major interest to Nikhil.(@adversarial_nik)